Path: utzoo!utgpu!news-server.csri.toronto.edu!cs.utexas.edu!usc!chaph.usc.edu!aludra.usc.edu
From: ajayshah@aludra.usc.edu (Ajay Shah)
Newsgroups: comp.ai.neural-nets
Subject: Re: Neural Network Training
Message-ID: <11143@chaph.usc.edu>
Date: 31 Jul 90 08:13:19 GMT
References: <6985@helios.TAMU.EDU>
Sender: news@chaph.usc.edu
Organization: University of Southern California, Los Angeles, CA
Lines: 28
Nntp-Posting-Host: aludra.usc.edu

In article <6985@helios.TAMU.EDU> vu2jok@cs.tamu.edu (Jogen K Pathak) writes:
>We are encountering problems while training the different paradigms , especially
>Back - Propagation paradigm. The training is very time consuming and tedious.
>Can anyone help to choose the training parameters' values that can
>reduce the training sessions. We are working in pattern classification of moderate size.e.g 100 input attributes.
>Any literature references also will be greatly appreciated.
>Jogen and Rajan.

I'd like to describe my experience.  I worked with a small sample
of 60 observations, with a discrete endogenous variable (takes
values 0/1/2) and 6 exogenous variables.  I spent one evening
trying to get a backprop network to produce sensible predictions
on a seperate set of 15 observations and failed miserably.

I used the BPS simulator program (a very nice program in case you
haven't tried it, except for the lack of offline operation) on a
386/387.  Each estimation took something like 15 minutes.  I
tried a diverse set of topologies and couldn't get anything which
performed well.

I don't know what I could be doing wrong.  Does anyone have ideas
on how to effectively converge upon backprop networks which work?

-- 
_______________________________________________________________________________
Ajay Shah, (213)747-9991, ajayshah@usc.edu
                              The more things change, the more they stay insane.
_______________________________________________________________________________